Better prediction of sub-cellular localization by combining evolutionary and structural information.

Area of Science:

• Computational biology
• Protein bioinformatics
• Structural biology

Background:

• Protein sub-cellular localization is crucial for function.
• Existing prediction methods rely on amino acid composition or known targeting motifs.
• Limitations exist in current methods, especially for proteins lacking known motifs or homology.

Purpose of the Study:

• To develop an accurate in silico method for predicting protein sub-cellular localization.
• To explore the utility of evolutionary information and protein structure for localization prediction.
• To address the performance gap observed with known structure databases.

Main Methods:

• Utilized evolutionary information from multiple sequence alignments.
• Incorporated aspects of protein structure into prediction models.
• Developed a hybrid system combining statistical rules and neural networks.
• Created separate prediction systems for proteins with known versus unknown structures.

Main Results:

• Achieved over 65% four-state accuracy, outperforming composition-based methods.
• Demonstrated highest accuracy for extracellular and nuclear proteins.
• Observed significant performance degradation for methods trained on SWISS-PROT when applied to PDB data.
• Successfully annotated eukaryotic proteins in the Protein Data Bank (PDB) using the developed system.

Conclusions:

• A novel computational approach integrating evolutionary and structural data enhances protein localization prediction accuracy.
• Distinct prediction models are necessary for proteins with known versus unknown structures.
• The developed system holds promise for target selection in structural genomics and proteomic annotation.